An intelligent interior design system based on 3D geometric modeling
By integrating artificial intelligence semantic networks, geometric constraint solvers, and physical simulation engines, real-time automatic correction of geometric data and real-time reconstruction of material lighting in interior design systems are achieved, solving the problems of geometric data discontinuity and offline rendering in existing technologies, and improving design efficiency and interactive experience.
Patent Information
- Application Number
- CN202611254041.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-18
- Publication Date
- 2026-11-17
AI Technical Summary
In existing interior design systems, geometric data gaps prevent the generated results from being directly used in industrial manufacturing. Physical legitimacy and ergonomic compliance verification are performed as a serial manual process. Offline rendering of materials and lighting results in low design efficiency and a poor parametric interactive design experience.
An intelligent interior design system based on 3D geometric modeling is adopted. By deeply integrating artificial intelligence semantic network, geometric constraint solver, physical simulation engine and differentiable rendering pipeline, it can realize real-time automatic correction of physical legality and ergonomic compliance constraints, as well as real-time reconstruction of materials and lighting after geometric shape modification.
The generated 3D model has a complete boundary representation topology, with dimensional accuracy down to the millimeter level. It supports CNC machining, and physical and ergonomic constraints drive layout solving in real time. The compliance time for high-density and complex spatial layouts is greatly shortened. After geometric modification, the material and lighting re-rendering latency is reduced to the real-time level, allowing designers to perform smooth parametric interactive design.
Smart Images

Figure CN123116394A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer-aided design and 3D geometric modeling engine technology, specifically involving an intelligent interior design system that deeply integrates artificial intelligence semantic prediction network, industrial-grade boundary representation geometric constraint solver, physical simulation engine and differentiable rendering pipeline. Background Technology
[0002] The digital workflow of interior design has long relied on a composite solution that combines traditional modeling engines with AI-powered rendering post-processing. A typical workflow is as follows: designers manually complete the 3D model of the interior space in traditional modeling software such as SketchUp or Autodesk Revit. Then, they stylize the 2D rendering of the scene using diffusion models. Next, the designers manually check and correct issues such as collisions and circulation according to design specifications. Finally, they perform offline rendering of lighting and materials in independent rendering software such as V-Ray or Lumion to output the renderings.
[0003] This type of solution suffers from three fundamental technical flaws.
[0004] First, the geometric data fragmentation prevents the generated results from being directly used in industrial manufacturing. The geometric data output by the artificial intelligence module is an unstructured mesh or implicit field representation, which lacks an automatic conversion bridge with the precise boundary representation topology required by the computer-aided design system. Key engineering parameters such as wall verticality and furniture coplanar fit cannot meet the millimeter-level accuracy requirements. The design results must undergo time-consuming manual reconstruction before they can be delivered to CNC machining equipment.
[0005] Second, the verification of physical legality and ergonomic compliance is a serial manual process, which leads to deadlock in high-density layouts. The existing solution places the verification of physical legality after the artificial intelligence generation, and the designer checks them one by one manually. The two processes are independent of each other and there is no information feedback loop. In high-density and complex spaces, a layout adjustment often triggers a chain reaction. The number of iterations of manual verification, correction and re-verification increases exponentially, resulting in extremely low design efficiency and a high risk of error.
[0006] Third, the offline rendering of materials and lighting is completely decoupled from geometry editing, resulting in a very poor parametric design experience. The rendering calculation of lighting and materials is performed offline in separate software, with no real-time linkage channel between it and the geometry modeling engine. Designers have to wait for several minutes or even tens of minutes of offline rendering after each adjustment of the geometry. Real-time preview is not feasible, which seriously restricts the experience and efficiency of parametric interactive design. Summary of the Invention
[0007] The purpose of this invention is to provide an intelligent interior design system based on three-dimensional geometric modeling. By deeply integrating artificial intelligence semantic networks, geometric constraint solvers, physical simulation engines and differentiable rendering pipelines, geometric data discontinuities are eliminated, and real-time automatic correction of physical legality and ergonomic compliance constraints is achieved, as well as real-time physical correct reconstruction of materials and lighting after geometric shape modification.
[0008] The present invention adopts the following technical solution:
[0009] An intelligent interior design system based on three-dimensional geometric modeling includes a semantic-geometric dual-flow modeling kernel module, a real-time layout correction module for physical and ergonomic constraints, and a material and lighting joint inversion module guided by geometric normals. The three modules work together through a bidirectional constraint computation graph to form an end-to-end closed-loop design process.
[0010] The semantic-geometric dual-stream modeling kernel module encapsulates boundary representation geometric primitives into differentiable tensor nodes that support automatic differentiation. These nodes are directly connected to the parameter input of the geometric constraint solver along with the output layer of the AI semantic prediction network, sharing the same backpropagation computation graph. The node types of both computation graphs are fully compatible. The semantic prediction network, based on the Transformer architecture, receives natural language descriptions or planar sketches input by the user. After processing by a multimodal encoder, it outputs a room function semantic vector, which is then mapped to the furniture type probability distribution and the initial geometric parameter vector via a fully connected prediction head. The initial geometric parameter vector is substituted into the corresponding type of boundary representation parameterized template in the parameterized furniture template library to instantiate an initial boundary representation geometric entity. The system performs geometric constraint checks on the initial boundary representation geometric entity, calculates the geometric constraint loss including size constraint penalty terms and coplanar fit constraint penalty terms, and weights and sums it with the semantic consistency loss to construct a joint loss function. The gradient of the joint loss function with respect to the geometric parameters and semantic prediction network parameters is calculated using an automatic differentiation mechanism, and gradient descent is performed iteratively until the geometric constraint loss is lower than the convergence threshold. Finally, a boundary representation geometric entity that meets industrial accuracy requirements is generated.
[0011] The joint loss function is expressed as follows:
[0012] ;
[0013] in, For the joint loss function, For semantic consistency loss, For geometric constraint loss, represents the global weighting coefficients for geometric constraint loss.
[0014] The geometric constraint loss is calculated as follows:
[0015] ;
[0016] in, Number the furniture. For the first Predicted geometric parameters of each piece of furniture. and These are the lower and upper bounds allowed for the geometric parameters, respectively. For the first The coplanarity error between the bottom surface of the furniture and the floor surface and These are the size constraint penalty coefficient and the coplanar constraint penalty coefficient, respectively.
[0017] The real-time layout correction module for physical and ergonomic constraints formalizes ergonomic standards and physical gravity laws into virtual geometric constraint entities, embedding them into the layout solution process in the form of directed distance field penalty terms. Specifically, it includes: defining an active bounding box for each piece of furniture based on its functional type and national standard human body size data; constructing a directed distance field expansion constraint field for the indoor active path; defining rotational hinge parameters and constructing an opening and closing envelope directed distance field for furniture with movable parts through parameterized skeleton binding; merging the above-mentioned wall directed distance field, active bounding box directed distance field, path width constraint directed distance field, and opening and closing envelope directed distance field into a single spatial constraint field by taking the minimum value operation; using an octree sparse acceleration structure to support real-time query; the system calls the rigid body physics engine to perform collision detection; constructing a physical constraint penalty loss based on the collision detection results and the unified spatial constraint field; calculating the gradient of each furniture pose parameter through automatic differentiation and performing iterative updates in a Lie group parameterization manner until the physical constraint penalty loss is lower than the physical constraint convergence threshold; and outputting a physically valid and ergonomically compliant final layout scheme.
[0018] The physical constraint penalty loss consists of three sub-items: rigid body collision penalty, directional distance field constraint penalty, and gravity legitimacy penalty. The directional distance field constraint penalty integrates the degree to which each furniture surface sampling point intrudes into the constraint region, and the gravity legitimacy penalty penalizes any furniture bottom surface located below the floor plane.
[0019] The geometric normal prior-guided material and lighting joint inversion module accurately extracts the unit normal vector and precise depth value in the camera coordinate system for each visible surface point from the boundary representation geometry engine using analytical normal calculation capabilities. These are organized into a normal map and a depth map, which are then injected as fixed constants into the solution space during subsequent optimization. The indoor scene is initialized as a set of three-dimensional Gaussian ellipsoids, each carrying spatial parameters, physically based rendering material parameters, and local spherical harmonic function lighting coefficients. For each Gaussian ellipsoid, the precise geometric normal is queried, and the effective surface normal is calculated. The normal offset attenuation coefficient is strictly less than 1 to ensure that the normal offset does not deviate excessively from the geometric prior. The radiation contribution of each Gaussian body is calculated using the GGX microsurface bidirectional reflection distribution function model and accumulated to obtain pixel color through an alpha blending mechanism. A joint inversion loss function is constructed, which includes rendering reconstruction loss, normal consistency loss, and material smoothing regularization term. The physically based rendering material parameters are decoupled from the lighting through gradient optimization. When the designer modifies the geometric parameters, the system immediately recalculates the accurate normal map and depth map of the affected area. The convergent material parameters are used as the initial values for hot start. Only a small number of gradient update steps are performed on the spherical harmonic function lighting coefficients to reconstruct the physically correct ambient light radiation distribution in the new geometric space. The target rendering latency does not exceed 33 milliseconds.
[0020] The present invention can achieve the following technical effects:
[0021] The system outputs a 3D model with a complete boundary representation topology and millimeter-level dimensional accuracy. It can be directly connected to CNC machining equipment, eliminating the cost of manual reconstruction. Physical and ergonomic constraints drive layout solving in real time in the form of gradients. The layout compliance time for high-density and complex spaces is compressed from hours of manual verification to hundreds of milliseconds of automatic convergence. The re-rendering latency of materials and lighting after geometric modification is reduced to the real-time level, with a target frame rate of no less than 30 frames per second. It supports designers to perform parametric interactive design and real-time presentation to clients. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of the five-layer overall architecture of the system of the present invention;
[0023] Figure 2 This is a flowchart of the semantic geometry dual-stream modeling kernel of the present invention;
[0024] Figure 3 This is a schematic diagram of the complete end-to-end closed-loop design of the present invention. Detailed Implementation
[0025] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.
[0026] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0027] System Overall Architecture
[0028] The intelligent interior design system of this invention consists of five core layers, such as... Figure 1 As shown, the bottom layer is the geometry kernel layer, which includes a boundary representation topology engine (based on the OpenCASCADE open-source kernel or a custom WebGPU kernel) and an automatic differential operator encapsulation module. The second layer is the bidirectional constraint calculation layer, the third layer is the physical and ergonomic constraint layer, the fourth layer is the material and lighting coordination layer, and the top layer is the user interaction layer, which provides a designer operation interface, an engineering drawing output interface, and an augmented reality roaming interface. Users can preview interior design results in a real-time immersive way through the augmented reality roaming interface, supporting extended reality (XR) interactive experience.
[0029] Specific implementation of the semantic geometry dual-stream modeling kernel module
[0030] The goal of this module is to build a modeling kernel that unifies artificial intelligence semantic prediction and industrial-grade geometric constraint solving within the same differentiable computation framework.
[0031] The system encapsulates each boundary representation primitive (vertex coordinates, edge parameters, and surface equation coefficients) in the geometric kernel into a differentiable tensor node that supports automatic differentiation. This is fully compatible with the computational graph node type of deep learning frameworks. The AI semantic prediction network is based on the Transformer architecture, and its output layer is directly connected to the parameter input of the geometric constraint solver. Both share the same backpropagation computational graph, forming an end-to-end differentiable system that fuses semantic and geometric streams.
[0032] Step 1: Semantic Feature Extraction and Initial Geometric Parameter Prediction
[0033] The natural language description (e.g., a Nordic-style four-person dining room) or floor plan sketch input by the designer is processed by a multimodal encoder to output a semantic vector of room function. ,in For the semantic embedding dimension, the semantic vectors are mapped to the furniture type probability distributions through a fully connected prediction head. and the initial geometric parameter vectors of various types of furniture ,in Indicates the first Furniture of this type These are the predicted values of the corresponding initial geometric parameters. For the geometric parameter dimension, the initial geometric parameter vector contains all the dimensional parameters required for the parameterized template representing the boundary of this type of furniture, including length, width, height, corner radius, etc.
[0034] Step 2: Instantiation of the boundary representation parameterized template
[0035] The system maintains a standardized parametric furniture template library. Each type of furniture in the library is stored as a parametric geometric template that conforms to the boundary representation of the topology. The template defines the complete vertex-edge-face topological relationships. The output from step one... Substitute the corresponding furniture type The boundary represents a parameterized template, generating the initial boundary representation of the geometric entity. :
[0036] ;
[0037] in, For the first The initial boundary of each piece of furniture represents a geometric entity. The boundary represents the template instantiation function. The type is Parametric geometric templates.
[0038] Step 3: Geometric Constraint Violation Detection and Penalty Loss Calculation
[0039] Perform geometric constraint checks on the initial boundary representation geometric entities generated in step two, including the following sub-items:
[0040] Dimensional constraint verification: Verify whether each geometric parameter meets the preset dimensional range constraints. ,in For the first The first piece of furniture One geometric parameter component, and These are the allowed lower and upper bounds for this parameter, respectively.
[0041] Topological consistency constraint check: Verify Euler's formula for the boundary representation entity Whether it is valid, among which , , , , These are the number of vertices, edges, faces, shells, and holes, respectively.
[0042] Coplanar fit constraint test: Verify the coplanarity error between the bottom surface of the furniture and the floor surface. Does it meet the requirements? ,in The maximum permissible geometric error threshold is set to 1 mm.
[0043] Taking all the above constraints into account, calculate the geometric constraint penalty loss:
[0044] ;
[0045] in, , These are the size constraint penalty coefficient and the coplanar constraint penalty coefficient, respectively, both of which are hyperparameters.
[0046] Step 4: Calculation of semantic constraint loss
[0047] To ensure that the result after geometric correction still aligns with the user's semantic meaning Figure 1 To calculate the semantic consistency loss:
[0048] ;
[0049] in, For the user's expectations True value tags for each furniture type For semantic vectors Predict the probability of the correct furniture type under given conditions. For the current geometric entity Extracted geometric semantic feature vectors, For the first The target semantic feature vector corresponding to each piece of furniture This represents the L2 norm.
[0050] Step 5: Construction of Joint Loss Function and Backpropagation of Gradients
[0051] Construct the total loss function:
[0052] ;
[0053] in, These are the global weighting coefficients for the geometric constraint loss, used to adjust the balance between semantic consistency and geometric accuracy.
[0054] Through automatic differentiation mechanism Regarding geometric parameters and semantic prediction network parameters Calculate the gradient and update iteratively using gradient descent:
[0055] ;
[0056] ;
[0057] in, For iteration rounds, and These are the learning rates for geometric parameters and semantic network parameters, respectively.
[0058] Step Six: Iterative Convergence Determination and Boundary Representation Entity Finalization
[0059] Repeat steps two through five until the convergence condition is met. ,in The geometric constraint loss convergence threshold is used. After convergence, the final parameters are... Substitute the parameterized template for the boundary representation to generate the final boundary representation geometry that meets industrial precision requirements. , and store in the system's geometric database.
[0060] Specific implementation of the real-time layout correction module based on physical and ergonomic constraints
[0061] This module formalizes the ergonomic standards and physical gravity laws of interior design into virtual geometric constraint entities, which are then embedded into the layout solution process in the form of a directed distance field penalty term.
[0062] Ergonomic standards are formalized into three types of virtual geometric constraint entities.
[0063] Active bounding box: For each piece of furniture, an axis-aligned bounding box larger than the geometric entity itself is defined according to its functional type. The expansion amount is determined based on the human body size data in the national standard GB / T10000-2023 (for example, a double bed is expanded by 600 mm on each side as a passage for getting up and lying down). The active bounding box is modeled as a soft exclusive area in the space. If the active bounding box of other furniture overlaps with it, a penalty will be incurred.
[0064] Minimum width constraint field for circulation routes: The indoor main circulation route is modeled as a one-dimensional manifold, and a directed distance field expansion is performed on it. The expansion radius is taken as the standard value of 600 mm, which is the minimum width of a single passage. The intrusion of all furniture in this constraint field is directly converted into penalty energy.
[0065] Cabinet door and drawer opening and closing envelope: For furniture with movable parts, the rotational hinge parameters for its opening and closing motion are pre-defined through parametric skeletal binding, including the hinge axis position. Rotation axis direction Maximum opening angle And an open and closed envelope oriented distance field is constructed using the outer envelope surface of the rotating swept body.
[0066] Construction of a unified spatial constraint field:
[0067] ;
[0068] in, Let the coordinates be any sampling point in space. For the directional distance field of the wall, For the first The directed distance field of a furniture activity enclosure box A directed distance field constrained by the width of the moving line. For the first The directed distance field of the opening and closing envelope of a piece of furniture. The minimum value among the subfields is taken to realize Boolean merging of the constraint fields, and the directed distance field adopts an octree acceleration structure for sparse storage and fast query.
[0069] The physical constraint penalty loss consists of the following three sub-items:
[0070] ;
[0071] in, For rigid body collision penalty terms, the penetration depth vector for all collision pairs. Find the sum of squares. This indicates the furniture pair that collided. For each furniture surface, there is a penalty term for the directed distance field constraint. The upsampling point encroaches on the constraint region at the transformed position (i.e. Integral with respect to the degree of ), For the first The positional changes of each piece of furniture This is a penalty item for the legality of gravity. For the first The lowest point of the base of the furniture in the world coordinate system coordinate, This is a gravity penalty factor that penalizes any furniture whose bottom surface is below the floor level.
[0072] Pose update uses Lie group parameterization to ensure manifold continuity during rotation:
[0073] ;
[0074] in, For Lie group index mapping, Update stride length for pose. For the first During the next iteration about The Lie algebra gradient is updated, and collision detection and penalty gradient calculation are repeated until... ( (Assuming a physical constraint convergence threshold), output the final compliant layout scheme. .
[0075] Specific implementation of the material and lighting joint inversion module guided by prior geometry normals
[0076] This module constructs a geometry, material, and lighting co-processing layer within the parametric mesh engine.
[0077] From the boundary representation geometry engine, the unit normal vector of each visible surface point is accurately extracted using the engine's analytical normal calculation capability. and precise depth values in camera coordinate system The organization is a normal diagram. and depth map ,in , These are the height and width of the rendered image (in pixels), respectively. In subsequent optimization, these two prior values are treated as fixed constants and do not participate in gradient updates.
[0078] Initialize the indoor scene as a set of three-dimensional Gaussian ellipsoids. ,in For the total number of Gaussian bodies, each Gaussian body Includes the following optimizable parameters: Spatial parameters including center position Covariance Matrix (by scaling vectors) and rotation quaternions (Decomposition to ensure semi-positive definiteness) Physically based rendering material parameters include primary color vectors. Roughness scalar Metallic scalar Normal offset vector The illumination parameters are the local spherical harmonic coefficients carried by each Gaussian body. (A total of 9 sets of coefficients of the 3rd order) are used to express the distribution of ambient light radiation at this location.
[0079] For each Gaussian body Query its center location In the normal diagram Precise normal corresponding to the projection of the mid-range camera Calculate the effective surface normal:
[0080] ;
[0081] in, This indicates that the vector has been normalized to a unit length. The normal offset attenuation coefficient and This is used to ensure that the normal offset does not deviate too much from the geometric prior. This is the effective normal used for the final calculation of the bidirectional reflection distribution function.
[0082] For each camera ray in the scene, the radiation contribution of each Gaussian body is calculated using the GGX microsurface bidirectional reflection distribution function model. The contributions of all Gaussian bodies on the ray are then weighted and accumulated using the alpha blending mechanism of 3D Gaussian sputtering to obtain the final rendered image. .
[0083] The joint inversion loss function is constructed as follows:
[0084] ;
[0085] in, For rendering images and reference images L1 reconstruction loss, The loss is due to the consistency between the effective normal and the geometric prior normal. For the corresponding weighting coefficients, For the total variational smoothing regularization term of the material parameters of adjacent Gaussian volumes, These are the corresponding weighting coefficients used to suppress high-frequency noise in the material parameters. This represents the total variational operator, processed by the Adam optimizer. Gradient descent iterations are performed on all optimizable parameters, and once convergence is achieved, the physically based rendering material parameters can be decoupled from the lighting.
[0086] When a designer modifies the boundary representation geometry parameters of a piece of furniture, the system executes the following real-time re-rendering process: The geometry engine immediately recalculates the accurate normal map and depth map of the affected area, injects the new normal prior into the effective normal calculation step, updates the effective normals of each Gaussian volume, uses the converged material parameters as the initial values for hot start, performs 5 to 10 gradient update steps only on the spherical harmonic function lighting coefficients, reconstructs the physically correct ambient light radiation distribution in the new geometric space, and completes the final rendering using the GPU hardware-accelerated ray tracing unit, achieving a real-time visual output update with a target latency of no more than 33 milliseconds.
[0087] A specific embodiment
[0088] The complete workflow of the system is illustrated using the living room design of a small apartment (45 square meters of building area) as an example.
[0089] The designer inputs a modern minimalist style in the user interface, describing a living room that can accommodate four people, requiring a TV area and a lounge sofa area. The semantic-geometric dual-stream modeling kernel module extracts semantic vectors and predicts the probability distribution of furniture types, identifying furniture types such as TV cabinets, three-seater sofas, coffee tables, and two single chairs. It generates initial geometric parameter prediction values for each type of furniture and instantiates the corresponding boundary representation parameterized template. After about 200 gradient iterations, the geometric constraint loss converges to below the threshold, generating accurate boundary representation geometric entities for all furniture.
[0090] The real-time layout correction module for physical and ergonomic constraints receives the aforementioned set of boundary representation entities and initial pose parameters, constructs a unified spatial constraint field, calls the BulletPhysics engine to perform collision detection (in this embodiment, clipping is detected between the initial positions of the sofa and coffee table), calculates the physical constraint penalty loss and performs pose gradient descent. After approximately 150 pose iterations, the physical constraint penalty loss converges, and a compliant layout scheme is output that satisfies the requirement that the width of a single passageway is not less than 600 mm and that the cabinet door opening and closing space is unobstructed.
[0091] The geometric normal-guided material and lighting joint inversion module receives the compliant layout scheme, extracts the precise normal map and depth map, initializes the scene as approximately 50,000 three-dimensional Gaussian ellipsoids, performs joint inversion optimization, and decouples and outputs physically based rendering material parameters such as base color, roughness, and metallic feel of each piece of furniture. The designer then adjusts the height of the coffee table from 400 mm to 450 mm. The system completes incremental lighting reconstruction within 33 milliseconds and outputs the updated real-time rendering. After the designer confirms the final scheme, the system exports the precise boundary representation geometric entities of all furniture as engineering drawings, which can be directly connected to CNC machining equipment.
[0092] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.
Claims
1. An intelligent interior design system based on three-dimensional geometric modeling, characterized in that, include: The semantic-geometric dual-stream modeling kernel module encapsulates boundary representation geometric primitives into differentiable tensor nodes that support automatic differentiation. This node is directly connected to the geometric constraint solver parameter input end of the output layer of the AI semantic prediction network, sharing the same backpropagation computation graph. The node types of both computation graphs are fully compatible. The semantic prediction network receives natural language descriptions or planar sketches from the user, outputs a room function semantic vector, and maps it to a furniture type probability distribution and an initial geometric parameter vector via a prediction head. The initial geometric parameter vector is substituted into a boundary representation parameterized template of the corresponding type in the parameterized furniture template library, instantiating an initial boundary representation geometric entity. The system performs geometric constraint checks on the initial boundary representation geometric entity, calculates the geometric constraint loss, and weights it with the semantic consistency loss to construct a joint loss function. The gradient of the joint loss function with respect to the geometric parameters and semantic prediction network parameters is calculated using an automatic differentiation mechanism, and gradient descent is performed iteratively until the geometric constraint loss is below the convergence threshold. Finally, the system outputs a final boundary representation geometric entity that meets industrial accuracy requirements. The real-time layout correction module for physical and ergonomic constraints is used to formalize ergonomic standards and physical gravity laws into virtual geometric constraint entities, construct a unified spatial constraint oriented distance field, call the rigid body physics engine to perform collision detection, construct a physical constraint penalty loss based on the collision detection results and the unified spatial constraint oriented distance field, calculate the gradient of each furniture pose parameter through automatic differentiation and perform iterative updates in the form of Lie group parameterization until the physical constraint penalty loss is lower than the physical constraint convergence threshold, and output the final layout scheme that is physically legal and ergonomically compliant. The geometry-normal-prior-guided material and lighting joint inversion module is used to accurately extract normal maps and depth maps from the boundary representation geometry engine using analytical normal calculation capabilities. During the optimization process, these are injected into the solution space as fixed constants. The indoor scene is initialized as a set of three-dimensional Gaussian ellipsoids, each carrying spatial parameters, physically based rendering material parameters, and spherical harmonic function lighting coefficients. A joint inversion loss function is constructed, which includes rendering reconstruction loss, normal consistency loss, and material smoothing regularization term. Through gradient optimization, the physically based rendering material parameters are decoupled from the lighting. When the designer modifies the geometry parameters, the convergent material parameters are used as the initial values for hot start, and only a small number of gradient update steps are performed on the spherical harmonic function lighting coefficients to achieve real-time rendering output updates.
2. The system according to claim 1, characterized in that, In the semantic-geometric dual-stream modeling kernel module, the joint loss function is expressed as follows: ,in For semantic consistency loss, For geometric constraint loss, The global weighting coefficient for the geometric constraint loss includes a size constraint penalty term and a coplanar fit constraint penalty term. The size constraint penalty term applies a squared penalty to geometric parameters that exceed the upper and lower bounds of preset parameters. The coplanar fit constraint penalty term applies a squared penalty to the coplanar error between the bottom surface of the furniture and the floor surface. The coplanar error threshold is 1 mm.
3. The system according to claim 1, characterized in that, In the semantic-geometric dual-stream modeling kernel module, a topological consistency constraint check is also performed for each boundary representation geometric entity to verify the Euler formula of the boundary representation entity. Whether it is valid, among which , , , , These are the number of vertices, edges, faces, shells, and holes, respectively. Entities that do not satisfy Euler's formula are marked as topological violations and trigger a re-instantiation process.
4. The system according to claim 1, characterized in that, In the real-time layout correction module for physical and ergonomic constraints, the virtual geometric constraint entities include the following three categories: The movable enclosure box is defined for each piece of furniture according to its function type and the human body size data defined in the national standard GB / T10000-2023. The expansion amount is determined by the ergonomic clearance specifications corresponding to the function type of the furniture. The minimum width constraint field of the circulation path models the indoor active circulation path as a one-dimensional manifold, and performs a directed distance field expansion on it, with the expansion radius taking the standard value of 600 mm, the minimum width of a single passageway; The opening and closing envelope oriented distance field is defined for furniture with movable parts by parametric skeleton binding, which defines the parameters of the rotating hinge, including the hinge axis position, rotation axis direction and maximum opening and closing angle, and constructs the opening and closing envelope oriented distance field with the outer envelope surface of the rotating sweep body.
5. The system according to claim 4, characterized in that, The unified spatial constraint directed distance field is constructed by merging the directed distance fields of walls, active bounding boxes, movement width constraint, and open / closed envelope by taking the minimum value point by point, and is stored and retrieved using an octree sparse acceleration structure.
6. The system according to claim 1, characterized in that, In the real-time layout correction module for physical and ergonomic constraints, the physical constraint penalty loss is composed of a weighted sum of three parts: a rigid body collision penalty term, a directed distance field constraint penalty term, and a gravity legitimacy penalty term. The rigid body collision penalty term is the sum of squares of the penetration depth vectors of all collision pairs. The directed distance field constraint penalty term is the integral of the degree to which each sampling point on the furniture surface intrudes into the negative region of the directed distance field of the unified spatial constraint. The gravity legitimacy penalty term applies a squared penalty to the case where the bottom surface of the furniture is lower than the floor plane. The pose update is performed through Lie group exponential mapping to ensure the continuity of the rotating manifold.
7. The system according to claim 1, characterized in that, In the material and lighting joint inversion module guided by the prior knowledge of geometric normals, the effective surface normal of each Gaussian body is calculated as a normalized weighted sum of the geometric normal and the normal offset vector. The normal offset attenuation coefficient is in an open interval between 0 and 1 and is much less than 1 to ensure that the normal offset does not deviate too much from the geometric prior. The normal consistency loss in the joint inversion loss function applies a squared penalty to the deviation between the effective normal and the geometric prior normal, thus constraining the solution space of material decoupling to a local neighborhood near the geometric normal.
8. The system according to claim 1, characterized in that, When the designer modifies the geometric parameters representing the furniture boundary, the material and lighting joint inversion module guided by the geometric normal prior executes an incremental re-rendering process: the geometry engine immediately recalculates the accurate normal map and depth map of the affected area, injects the new normal prior into the effective normal calculation steps, and performs only 5 to 10 gradient update steps on the spherical harmonic function lighting coefficients with the converged material parameters as the initial values for hot start, with the target rendering latency not exceeding 33 milliseconds.
9. The system according to claim 1, characterized in that, The system also includes a user interaction layer, providing a designer operation interface, an engineering drawing output interface, and an augmented reality roaming interface. After the designer confirms the final layout plan, the system exports the precise boundary geometric entities of all furniture into engineering drawings that can be directly connected to CNC machining equipment. The augmented reality roaming interface supports extended reality interactive indoor scene roaming preview output.
10. The system according to claim 1, characterized in that, The boundary representation geometry engine is implemented based on the OpenCASCADE open-source industrial-grade geometry modeling kernel, or based on a custom WebGPU geometry kernel. Each type of furniture in the parametric furniture template library is stored as a parametric geometry template that conforms to the boundary representation topology. The template defines a complete vertex-edge-face topological relationship. The system outputs a 3D model with a dimensional accuracy of millimeters.
